← projects echo-mind
active

echo-mind

A RAG-powered chatbot that answers questions about my portfolio, built on FastAPI, Lambda, and Claude

// stack
  • #Python
  • #FastAPI
  • #AWS
  • #Terraform
  • #LangChain
  • #Anthropic
  • #Docker
[src] view source
echo-mind.md

Overview

echo-mind is a retrieval-augmented generation chatbot that knows about my work. Ask it about a project, a blog post, or anything in my portfolio and it retrieves the relevant content and answers grounded in what’s actually there, not what an LLM guesses.

The architecture is a FastAPI app running on AWS Lambda via container image, sitting behind API Gateway and a Cloudflare-managed custom domain. The full stack is provisioned with Terraform.

Status

The project is being built in four phases. Phases 1 and 1.5 are complete.

PhaseWhatStatus
1FastAPI on Lambda, API Gateway, ECR, S3 scaffoldingdone
1.5Custom domain: Cloudflare DNS + ACM certificatedone
2Document ingestion, chunking, embeddings, FAISS indexin progress
3RAG chain: retriever + Claude for grounded answersupcoming
4Chat memory, token streaming, frontend integrationupcoming

Architecture

Browser
  -> Cloudflare DNS (api.dachava.dev)
  -> API Gateway v2 HTTP API
  -> Lambda (container image from ECR)
  -> Mangum (Lambda event to ASGI)
  -> FastAPI

At query time (Phase 3 and beyond):

user question
  -> embed query (text-embedding-3-small)
  -> FAISS similarity search over portfolio chunks
  -> top-k chunks injected into Claude prompt
  -> grounded answer streamed back

What I Learned Building It

Phase 1 forced a clear understanding of why Lambda needs an ASGI adapter (Mangum bridges the Lambda event format and FastAPI’s HTTP expectations), why container images beat zip files for Python projects with real dependencies, and how Terraform remote state with S3 locking works in practice.

Phase 1.5 introduced the multi-vendor Terraform problem: coordinating ACM certificate validation across AWS and Cloudflare in a single terraform apply. The key insight was using the Cloudflare provider to create the DNS validation record before aws_acm_certificate_validation starts polling, eliminating the manual step that makes this workflow painful.

Both phases are documented in detail in the blog posts below.

← back to projects